B Cell-Related Prognostic Model in Gastric Cancer via Integration of Single-Cell Sequencing and Transcriptome Sequencing Data

Technical Summary: B Cell-Related Prognostic Model in Gastric Cancer via Integration of Single-Cell and Transcriptome Sequencing Data

Problem Statement
Gastric cancer (GC) remains a leading cause of cancer-related mortality globally, often diagnosed at advanced stages where chemotherapy yields limited efficacy. While immunotherapy offers a promising alternative, patient responsiveness varies significantly, largely governed by the functional states of immune cells within the tumor microenvironment (TME). Specifically, tumor-infiltrating B cells (TIBs) are recognized as critical modulators of antitumor immunity, yet their specific transcriptomic signatures and prognostic value remain under-explored. A major computational challenge lies in selecting stable, generalizable prognostic markers from high-dimensional genomic data. Existing models often rely solely on bulk RNA sequencing, which captures average tissue expression but fails to resolve intercellular heterogeneity, or single-cell sequencing, which lacks population-level context. This study addresses the need for a robust biomarker discovery framework that integrates these data modalities to construct a reliable B cell-related prognostic model for GC.

Methodology
The study employed a multi-step integrative bioinformatics workflow combining single-cell RNA sequencing (scRNA-seq) and bulk RNA sequencing (bulk RNA-seq) data:

1. Data Acquisition: Researchers retrieved scRNA-seq data (GSE183904) and bulk RNA-seq data (GSE63089) from the Gene Expression Omnibus (GEO), along with transcriptome and clinical data from The Cancer Genome Atlas (TCGA). Four independent GEO datasets (GSE51105, GSE15459, GSE22377, GSE62254) served as external validation cohorts.
2. Single-Cell Analysis: Using the Seurat package, scRNA-seq data underwent quality control (filtering cells with 50–5,000 genes and <5% mitochondrial reads), normalization, and batch effect correction (Harmony). Dimensionality reduction (PCA and UMAP) and clustering identified nine cell subpopulations. Differentially expressed genes (DEGs) specific to B cells were identified using the FindMarkers function (logFC > 1, adj. p < 0.05).
3. Bulk Transcriptome Analysis: DEGs between GC tumor and adjacent normal tissues were identified from bulk data using the limma package (|log₂FC| > 1, p < 0.05).
4. Gene Selection and Model Construction: The intersection of B-cell-related DEGs from scRNA-seq and bulk DEGs yielded 76 overlapping genes. From these, 35 genes with complete annotation in the TCGA cohort were subjected to multivariate Cox stepwise regression. This process refined the signature to two core hub genes: CTHRC1 and CD44.
5. Risk Score Calculation: A prognostic risk score was calculated for each patient using the formula: Risk Score = (Expression of CTHRC1 × Coefficient) + (Expression of CD44 × Coefficient).
6. Validation and Evaluation: Patients were stratified into high- and low-risk groups based on the median risk score. Model performance was evaluated using Kaplan-Meier (K-M) survival curves, time-dependent Receiver Operating Characteristic (ROC) curves, and univariate Cox regression to assess independence from clinical factors (age, sex, TNM stage). Furthermore, the discriminative capacity was tested using nine machine learning algorithms (KNN, SVC, Logistic Regression, Decision Tree, Random Forest, MLP, AdaBoost, Gradient Boosting, and Bagging).

Key Results

• Gene Identification: The study successfully identified CTHRC1 and CD44 as the two core hub genes associated with B cell infiltration and GC prognosis.
• Prognostic Performance: In the internal TCGA cohort, the high-risk group exhibited significantly shorter overall survival compared to the low-risk group (p = 0.017). The risk score was confirmed as an independent prognostic factor (p < 0.001).
• Predictive Accuracy: The model demonstrated favorable predictive accuracy with Area Under the Curve (AUC) values of 0.550 (1-year), 0.627 (3-year), and 0.874 (5-year) in the TCGA cohort.
• Robustness: The model's generalizability was confirmed in external validation cohorts, where high-risk patients consistently showed worse prognosis (p = 0.0014).
• Machine Learning Validation: All nine machine learning classifiers achieved AUC values exceeding 0.7, indicating that the two-gene signature consistently captures the underlying risk stratification signal regardless of the algorithmic approach.

Significance and Claims
The authors claim that by integrating scRNA-seq and bulk RNA-seq data, this study bridges cellular-resolution insights with population-level expression patterns to identify B cell-related core genes. The constructed two-gene prognostic risk model (based on CTHRC1 and CD44) provides a reliable reference for the prognostic evaluation of gastric cancer.

The study highlights the biological relevance of the selected genes:

• CTHRC1: A secreted extracellular matrix protein overexpressed in GC, implicated in microenvironment remodeling, migration, and immune evasion.
• CD44: A canonical cancer stem cell marker associated with invasive fronts, recurrence, and drug resistance.

The paper concludes that this B cell immunity-associated risk model offers a theoretical foundation and analytical framework for clinical prognostic assessment and the formulation of immunotherapeutic strategies in gastric cancer, without overclaiming immediate clinical deployment.